Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Monitoring the key components of a biochemical process can be done through online and offline measurements. However, offline measurements are not applicable for online control and available online probes are often unreliable and expensive. The use of a soft sensor can reduce the needed number of online probes by combining sensor data and model predictions. A design of a soft sensor is proposed in this thesis for monitoring a lab-scale aerobic cultivation of E. coli. The extended Kalman filter (EKF) is used for the estimation of optical density (OD), glucose, acetate and dissolved oxygen (DO). The EKF uses a non-linear model for the prediction and online (noisy) measurements for the correction. The employed model captures fermentative and oxidative growth on glucose and the used online measurements are OD and DO. First, the model parameters are identified based on batch and fed-batch data of performed experiments. A confidence interval of the derived parameter estimates is conducted based on the Fisher information matrix. Next, the EKF is fine-tuned and an observability analysis is performed. Two additional augmented EKF's are used and fine-tuned. The first augmented EKF estimates the time-varying oxygen mass transfer coefficient in addition and the second EKF is augmented with the maximum glucose uptake rate to deal with model uncertainties. Lastly, the fine-tuned EKF's are validated. The identified model parameters based on combined batch and fed-batch data showed a low confidence in the glucose saturation constant due to overfeeding of glucose. The estimated acetate inhibition constant was lower compared to values found in literature. Through the condition number of an observability matrix, it is mathematically shown that the EKF cannot observe the real state for certain operating conditions. The EKF is not able to observe high glucose concentrations, as its effect does not show up in the online measurements. Therefore, glucose estimates of the EKF heavily rely on the model prediction. However, glucose estimates deviated from the offline measurements, due to model uncertainties. A high trust in the acetate model estimate was put to limit strong corrections made on the acetate estimate. The fine-tuned EKF was validated on new experimental data. Again, model uncertainties caused a mismatch between glucose estimates and offline measurements, but acetate was very well estimated. The augmented EKF estimated the oxygen mass transfer coefficient well. However, a good performance highly depends on correctly estimated oxygen uptake rates. The augmented EKF estimating the maximum glucose uptake rate showed an improved performance by dealing with model uncertainties.